Test-Time Instance-Specific Parameter Composition: A New Paradigm for Adaptive Generative Modeling
This work addresses the limitation of static generative models that cannot adapt to individual inputs, offering a method for per-instance specialization with minimal overhead.
Composer introduces a new paradigm for adaptive generative modeling by generating input-conditioned parameter adaptations at test time, which are injected into pretrained model weights. This approach improves performance across diverse generative models, including lightweight/quantized models, without fine-tuning or retraining.
Existing generative models, such as diffusion and auto-regressive networks, are inherently static, relying on a fixed set of pretrained parameters to handle all inputs. In contrast, humans flexibly adapt their internal generative representations to each perceptual or imaginative context. Inspired by this capability, we introduce Composer, a new paradigm for adaptive generative modeling based on test-time instance-specific parameter composition. Composer generates input-conditioned parameter adaptations at inference time, which are injected into the pretrained model's weights, enabling per-input specialization without fine-tuning or retraining. Adaptation occurs once prior to multi-step generation, yielding higher-quality, context-aware outputs with minimal computational and memory overhead. Experiments show that Composer substantially improves performance across diverse generative models and use cases, including lightweight/quantized models and test-time scaling. By leveraging input-aware parameter composition, Composer establishes a new paradigm for designing generative models that dynamically adapt to each input, moving beyond static parameterization.